Comment by nextos

The problem with deep learning right now is that there's not a great deal of methods outputting a nice estimation of uncertainty in predictions. But this will become less of an issue as they eventually converge with Bayesian methods, or at least if probability is introduced in a more principled way.